A compressor rolls off the line in Shunde every few seconds. For years, a worker squinted at each one for scratches, dents and misaligned parts until fatigue set in. Today a camera and a model do the first pass, and the human only steps in when something looks genuinely wrong.
This is not a pilot poster child. It is becoming the baseline.
Across Chinese manufacturing, AI visual inspection has moved from showcase to default faster than almost any other applied-AI use case on the shop floor.
The number that surprised even insiders
In 2025, the AI application rate in the quality inspection step of large-scale manufacturers hit 82.7% — the highest penetration of any industrial AI scenario, ahead of R&D, scheduling and equipment maintenance. The reason is unglamorous: inspection is repetitive, visual, and directly tied to money lost on returns and recalls.
Where it has landed:
- Baosteel (宝钢) runs machine-vision quality systems across steel production, from raw material to finished coil.
- Haier (海尔) deploys AI visual inspection at its Shanghe refrigerator plant under the COSMOPlat "Tianzhi" industrial large model, with reported inference accuracy above 96%.
- Weld-defect recognition in advanced smart-manufacturing projects reaches 97%+ accuracy on validated samples.
A 1.3-billion-parameter eye in Shunde
The deployment drawing the most attention is at GMCC (美芝), a Midea-group compressor unit, where China Unicom (中国联通) stood up what it calls the first "AI + visual large model" in the home-appliance industry, live on the production line at the Xingtan plant in August 2025.
The technical story is what makes it different from the old "one model per defect" approach:
- A Transformer-based architecture with 1.3 billion parameters (13亿参数), trained on more than 100 million industrial defect images.
- Few-shot learning: each defect type needs only 1–5 reference samples to adapt, slashing annotation time by over 80%.
- Training cycle shortened by over 80%, development cost down 66%, and a "zero-intrusion" appliance-style deployment that finishes full-line testing within 72 hours without stopping production.
For an older factory, that means intelligent inspection can be added without a tear-down — the model platform supports cross-line, cross-model reuse.
Why this is spreading beyond the giants
The earlier blocker was data and downtime: you needed thousands of labeled defect images and a line you could afford to pause. Large models flipped both constraints.
- Labeling a new defect used to take weeks of expert time; now it is a handful of samples.
- Edge and 5G-A setups push inference to the line side, so a camera can flag a flaw in milliseconds rather than shipping images to a distant server.
- Hardware — industrial cameras, edge boxes — keeps getting cheaper, which is what lets mid-size plants join, not just the Baosteels of the world.
Another quiet driver is the data flywheel. Every defect a line catches becomes a labeled example that improves the next model, so accuracy compounds instead of decaying. Plants that once feared "garbage in, garbage out" now treat inspection as a feedback loop feeding the same model family across sites.
IDC, via the AI Industry Alliance (AIIA), projected China's industrial intelligent-inspection market would reach US$958 million (about 6.2 billion RMB / HK$7.4 billion) by 2025, at a 28.5% compound growth rate from 2021 — a sign the category is treated as infrastructure, not experiment.
From showcase to baseline
The strategic prize is not one saved inspector but the "black-light factory" (黑灯工厂) — a line that runs with the lights off because no human needs to be there to watch it. AI inspection is the sense organ that makes that safe: it catches the batch defect at the source instead of after it ships. That is why provincial smart-manufacturing subsidies and the national "AI+" action both treat visual inspection as a priority layer rather than a nice-to-have.
The human is not gone
At the Yanfeng (延锋) plant in Shanghai's Jiading district, an AI target-detection system cut tailgate inspection time by 46 seconds per unit, freeing roughly 9.2 hours a day across two lines. The point was not to fire inspectors but to redirect them: humans handle the borderline calls, the model handles the obvious 90%.
Honest limitations
High reported accuracy (96–97%+) is measured on curated, in-distribution samples and will dip on new product variants or unfamiliar lighting. Models still need domain data to generalize across product lines, and false positives create re-work if humans aren't in the loop. For safety-critical parts, AI inspection is a filter, not a sign-off — regulatory and customer requirements still demand human or redundant verification.
What readers can do now
- If you run a production line, start with the highest-volume, most standardized part — that is where visual models pay back fastest and need the least custom data.
- Ask vendors for a 72-hour proof-of-concept on a live line, not a lab demo; deployment without stopping production is now realistic.
- Track defect escape rate, not just detection rate — the metric that actually moves your return and warranty cost is what the model misses.